PhysicalAI-Robotics-Kitchen-Sim-Demos

PhysicalAI-Robotics-Manipulation-Kitchen-Demos 是由 NVIDIA 发布的面向 Physical AI 与机器人操作学习的大规模机器人示范数据集，用于训练和评估具身智能系统在真实物理环境中的操作能力。该数据集包含约 600 小时人类遥操作示范数据，覆盖 316 类厨房操作任务，累计约 5.5 万条机器人操作轨迹，采集平台包括 Franka Panda 机械臂和 Omron 移动底座。数据聚焦厨房场景中的抓取、移动、开关柜门、物品整理等复杂物理交互任务，支持模仿学习（Imitation Learning）、视觉-语言-动作模型（VLA）及机器人策略学习。数据采用 LeRobot 标准格式，包含任务描述、机器人状态、动作轨迹、多视角视频等多模态信息，并遵循 CC BY 4.0 开源协议，适用于机械臂操作、家庭服务机器人及 Physical AI 研究。PhysicalAI-Robotics-Manipulation-Kitchen-Demos is a large-scale robotic demonstration dataset released by NVIDIA, tailored for Physical AI and robotic manipulation learning, and used to train and evaluate the operational capabilities of embodied intelligence systems in real physical environments. This dataset contains approximately 600 hours of human teleoperation demonstration data, covering 316 types of kitchen manipulation tasks, and accumulating a total of around 55,000 robotic manipulation trajectories. The data collection platforms include the Franka Panda robotic arm and the Omron mobile base. The dataset focuses on complex physical interaction tasks in kitchen scenarios, such as grasping, moving, opening and closing cabinet doors, and tidying up items. It supports research on Imitation Learning, Vision-Language-Action (VLA) models, and robotic policy learning. The data is stored in the LeRobot standard format, including multimodal information such as task descriptions, robot states, action trajectories, and multi-view videos. It adheres to the CC BY 4.0 open-source license, and is applicable to research on robotic manipulation, home service robots, and Physical AI.